Overview
Noema is Agent Space's own distilled model line. Models are served through ArcRouter — callers use ArcRouter model aliases and never hit the inference layer directly. The SDK is effectively the ArcRouter client configured with Noema-specific aliases.
The long-term goal: local Noema models at parity with frontier models for Grid task classes, removing the Anthropic dependency entirely for internal work.
Ratified decisions (NOEMA.0)
- D-N01 - Base model: Qwen2.5 14B (not Llama, not Mistral)
- D-N02 - Training data: owned frontier-pairs + permissive OSS blend
- D-N03 - Eval gate: approved. Evals must pass before a model version is promoted.
ArcRouter aliases
Noema models are registered in ArcRouter as named aliases. Use these in any Grid code — the router resolves to the best available serving node.
# Noema ArcRouter aliases (NOEMA.1+)
AR-D1 → noema-base # Qwen2.5-14B base instruction
AR-D2 → noema-base-moe # Qwen3.6-35B-A3B mixture-of-experts
AR-D3 → noema-tenant-fail-closed # tenant isolation enforcement
AR-D4 → noema-image-gen # FLUX.2 klein 4B (NOEMA.5)
AR-D5 → noema-image-edit # Qwen-Image (NOEMA.5)Quick start (via Brain SDK)
Route Noema calls through the Brain SDK like any other model. Use the AR-D alias as the model name.
import { BrainClient } from '@arcgrid/brain-client'
const brain = new BrainClient({ tenantKey: process.env.BRAIN_KEY })
const res = await brain.chat({
model: 'AR-D1',
messages: [{ role: 'user', content: 'Draft a project status update.' }]
})Adapters (NOEMA.3+)
Department and specialty LoRA adapters stack on top of the base model. They are registered in ArcHub and resolved by ArcRouter based on the calling tenant's department context. No code changes required to switch adapters — the router selects the best one.
Build status
NOEMA.0 (license, provenance, decisions) is ratified. NOEMA.1 (first serve: vLLM + 96 GB card + ArcRouter wired) depends on hardware procurement and CLOUDPROD phase. See the Noema Capability Matrix for full milestone status.